[prev in list] [next in list] [prev in thread] [next in thread] 

List:       xvid-devel
Subject:    Re: [XviD-devel] Other clip with Bframes, quality vs size
From:       Christoph Lampert <chl () math ! uni-bonn ! de>
Date:       2002-08-18 17:19:48
Message-ID: Pine.LNX.4.21.0208181905290.15624-100003 () lieb ! math ! uni-bonn ! de
[Download RAW message or body]


The last clip, this time Michael's favourite "Dogma" Clip. actually it's
rather interesting:

The first one is as before: bpp versus PSNR again.

You see the great gap between PSNR 46 and 52, which is in fact between 
Pframes' quant=3 and Pframes quant=4. This is most likely because of
prequantized input material. 
I guess most parts of the material was encoded at quantizer 3 before, so
encoding at quant=3 again gives almost perfect reconstruction. 

But my topic is PSNR<->bpp for B-frames, and because of the huge gap you
can't see anything, so there is a second plot (...lower.png) which shows
just the results for quant 4 and higher. 

There you see that sometimes Bframes are better and sometimes not. Great, 
what a statement! You cannot tell which max_bframes value would be best to
choose from this plot. 

So there is another plot, where not max_bframes gets different colors, but
bquant_ratio (Dogma-bquantratio.png). Here the colors mean:

Red+line: max_bframes=0 (so bquant doesn't matter)
 
green: bquant_ratio = 100    result: always worse then non-bframes
blue:  bquant_ratio = 150    result: always better than non-bframes
purple:bquant_ratio = 200    result: even better in some cases
turquoise:    ratio = 250    result: best, in many cases (!) 

ratio=80 was left out, because it's clearly inferior.

So, what parameters to choose for an arbitrary file? 

I don't have the faintest idea...

gruel



P.S. Be aware that these just the _average_ values. Of course 
P-frames quality is better than B-frames quality for ratio>100 and not
only average value, but standard deviation has to be taken into
account. I did those plots, too, but they are not help, either. 

-- 
Christoph H. Lampert chl@math,uni-bonn,de | Diese Signature wurde maschi-     
Beringstr. 6, Raum 14 Tel. (0228) 73-2948 | nell erstellt und bedarf
Sprechstunden: keine, aber meistens da    | keiner Unterschrift. AZ 27B-6 

["Dogma-quality-vs-size.png" (IMAGE/png)]
["Dogma-quality-vs-size-lower.png" (IMAGE/png)]
["Dogma-bquantratio.png" (IMAGE/png)]

[prev in list] [next in list] [prev in thread] [next in thread]